Maker-Taker Fee, Liquidity Competition, and High Frequency Trading *

Transcription

1 Maker-Taker Fee, Liquidity Competition, and High Frequency Trading * Yiping Lin a, Peter L. Swan b, and Frederick H. deb. Harris c, This Draft: February 1, 2017 Abstract This paper analyzes how a unilateral maker-taker fee reduction by Nasdaq affects market competition, liquidity, and high frequency trading. Holding the net exchange fee relatively constant, we find that: 1) contrary to much of the literature, the components of maker-taker fee change do matter; 2) the removal of the subsidy to non-marketable orders causes the flight of informed market orders to the remaining highest tax-subsidy venues; and 3) the flight of informed orders reduced Nasdaq s market share and pricing efficiency. By using the Nasdaq maker-taker fee experiment and difference-in-differences methodology, we find the reduced maker rebate lowers quote quality and the percentage of incoming orders routed to Nasdaq. However, in an only partially offsetting manner, it improves fill rate and speed of fill due to the reduced taker fee. Any improved relative position of a market in routing tables causes adverse selection costs to decline, together with information in the market orders, standard measures of market liquidity to increase, and liquidity supplier profits to decline. The maker-taker fee reduction did not result in a gain in share from off-exchange trading venues. As the fee and rebate decline, high frequency traders tend to switch from adding to removing liquidity. Keywords: Maker-Taker Fee; Market Fragmentation; Market Quality; High Frequency Trading JEL Classifications: G12, G14 * The views expressed herein are strictly those of the authors. Yiping Lin acknowledges the financial support from the CMCRC and Wake Forest University for facilitating his work in the United States. We thank Nasdaq, Inc. for providing data while Yiping Lin was visiting under an agreement between Nasdaq and the CMCRC. We also thank Avanidhar Subrahmanyam, Joakim Westerholm, Kumar Venkataraman and seminar participants at Australasian Finance & Banking Conference for helpful comments and suggestions. The views herein are not intended to represent the views of Nasdaq, Inc., its employees, or directors. Any errors or omissions are the responsibility of the authors alone. a School of Banking and Finance, UNSW Business School, UNSW, Sydney NSW 2052 Australia; yiping.lin@unsw.edu.au. b School of Banking and Finance, UNSW Business School, UNSW, Sydney NSW 2052 Australia; peter.swan@unsw.edu.au. c School of Business, Wake Forest University, Winston-Salem, NC 27106, United States; harrisf@wfu.edu. 1

2 I. Introduction The importance of the components of an exchange s maker-taker fee remains unclear. Previous studies such as Angel, Harris, and Spatt (2011), Colliard and Foucault (2012), and Malinova and Park (2015) state that only the net fee change matters and that the maker-taker fee breakdown is irrelevant since the tax on market orders appears to offset the subsidy to limit orders. In fact, Foucault, Kadan, and Kandel (2013) go one step further to claim that trading volume may increase or decrease (depending on the model parameters), even in the absence of a change in the net total fee, because a fixed tick size prevents prices from neutralizing the effect of the maker rebate. This potential cancelling out is in appearance only, because maker-taker fees are normally mistakenly treated as a tax on market orders and offsetting subsidy to limit orders. Not so. The exchanges that devised maker-taker fees were fully aware of potential offset. Hence they limit the subsidy to non-marketable orders which means, in effect, that only orders with delayed execution that increase the depth of the limit order market receive the subsidy. If the tax and subsidy were fully offsetting, depth would be unaltered and this could only be the case if marketable limit orders, as well as non-marketable limit orders, received the subsidy. This ongoing debate on the exchange fee structure in U.S. markets led NASDAQ to experiment with lowering both maker rebate and taker fee by an identical amount for a select group of securities. While the Nasdaq 3 experiment was motivated by competition from dark and off-exchange venues, the real benefit of the experiment lies it its ability to reveal for the first time why maker-taker fees are of critical importance for almost every aspect of market quality and why tax-subsidy schemes are fundamental with no prospect of ever cancelling out. One way to think about what really happens is that the best bid and ask are ideal for informationless liquidity trades designed to minimize consumption volatility risk but these do not attract subsidy as these are, in effect, marketable orders. Orders in the limit order book queue further away that enhance depth and facing either delay or non-execution risk are the both the target of the subsidy and provide limit order places some element of protection from informed traders who move price substantially. It is clear then that a subsidy to market depth can only benefit one class of trader, namely informed traders, and, 3 Over the past decade, hundreds of exchange-listed securities have traded more volume in off-exchange markets than on exchange markets. In response to claims that public markets are too expensive to trade, Nasdaq wanted to know whether access fees may be discouraging the use of public markets 3 (Nasdaq s SEC, 2014). 2

3 moreover, the same class of trader is the only one willing to pay the high tax on market orders as only they are effectively able to benefit from the greater depth of the limit order book. Impatient limit order providers in the now much deeper book are now willing to become the counterparty to the informed traders now dominating the taker market in the presence of a higher taker fee. Hence the maker-taker fee structure is cleverly designed to cancel out for only one trader type the informed. In this paper, we show that the maker-taker fee structure employed by Nasdaq has highly desirable features in that it improves price discovery and market efficiency while at the same time maximizing Nasdaq s trading volume and market share. By temporarily unwinding these highly desirable features of the Nasdaq exchange for a limited number of stocks and a limited time-frame, Nasdaq was able to reveal for the first time why its fee structure was so successful. It is hard for profitorientated exchanges to provide for what is essentially a public good namely efficient price discovery but the taker-maker fee structure has evolved as a natural counterpart to competition in so-called fragmented markets. In this study, we use the difference-in-differences (DID) method to examine the impact of the NASDAQ exchange access fee change on market competition, liquidity, and high frequency trading (HFT) behavior. Specifically, we first test whether the components of maker-taker fee changes matter when the net exchange fee remains relatively constant and find that they are of critical importance. Second, we test whether reducing the exchange maker-taker fee attracts liquidity from off-exchange venues to find that it does not. This is because of the public-good aspect of maker-taker fees alluded to above means that alterations to subsidy-tax arrangements on price discovery do not affect off-lit market exchanges. Third, we extend the purpose of the Nasdaq pilot to test how HFT behavior changes in response to the maker-taker fee change. Since we show that HFT switches from adding to removing liquidity as the taxsubsidy is removed, this novel finding indicates that HFT is a distinct trader-type, neither informed nor uninformed, and has a substantial niche regardless of tax-subsidy schemes. The Nasdaq access fee reduction pilot offers an informative natural experiment. Nasdaq charge taker fee for removing liquidity by submitting marketable orders, and provide maker rebate for adding liquidity by submitting non-marketable orders (i.e., limit orders which cannot be executed immediately). On February 2, 2015, Nasdaq implemented a maker-taker fee pilot for 14 traded stocks on Nasdaq where the taker fee was lowered to 5 cents per 100 shares (CPS) from 30 CPS to remove displayed liquidity; the 3

4 maker rebate for adding displayed liquidity was lowered to 4 CPS from an indicative 4 29 CPS. For simplicity, the indicative rebate for the PilotOff period is summarized in Table 1. The pilot ended on May 31, 2015, when the fee reverted to its pre-pilot level. <Insert Table 1 here> The International Organization of Securities Commissions defines maker-taker fees as a pricing model whereby the maker of liquidity, or passive [limit] order, is paid a rebate and the taker of liquidity, or aggressive [market] order, is charged a fee. Maker-taker fees for protected quotes in the equities markets are bound by Rule 610 of Regulation NMS 5, which caps fees at 30 CPS traded. Maker rebates aim to both improve liquidity, by rewarding its provision, and increase trading volume. The maker-taker payment model originated with electronic trading venues in the late 1990s (Harris, 2013). In 1997, the Island electronic communication network (ECN) was among the first markets to adopt maker-taker fees, which attract order flow through liquidity rebates. These rebates provide traders with an additional source of income other than the bid-ask spread, incentivizing liquidity providers to post more competitive quotes to attract order flows from other markets. As a result, Island s market share of reported Nasdaq trades increased from approximately 3% in 1997 to almost 13% in 1999 (Cardella, Hao, and Kalcheva, 2015). Other alternative trading systems (ATSs) soon followed Island s fee model to attract liquidity and order flows from lit exchanges (SEC, 2015a). <Insert Table 2 here> In response to the ATS competition, many exchanges adopted maker-taker fees of their own. Over the past decade, the maker-taker pricing model has thus gained widespread adoption in the U.S. equities 4 The maker rebate scheme is much more complicated (see Further, see ITG Takeaways from the NASDAQ Pilot Program report, available at 5 If the price of a protected quotation is less than $1, the fee cannot exceed 0.3% of the quotation price. See SEC Rule

5 market (see Table 2), rewarding liquidity suppliers and charging liquidity demanders. Only three U.S. exchanges (BATS-Y, EDGA, Nasdaq/BX) adopt the inverted fee (taker-maker) model, offering rebates to take liquidity accompanied by a higher fee in order to add liquidity. Angel, Harris, and Spatt (2011) argue that introducing a maker rebate financed by a taker fee should have no effect, because prices in competitive markets adjust by the rebate amount. However, this argument assumes that the information content in order flows is the same across venues and breaks down when different venues have different fee structures. Colliard and Foucault (2012) formalize Angel, Harris, and Spatt s intuition and prove, without relying on perfect competition, that in the absence of frictions only changes in the net total fee retained by the exchange affect liquidity and trading volume. Foucault, Kadan, and Kandel (2013), by contrast, show that trading volume may increase or decrease (depending on the model parameters), even in the absence of a change in the net total fee, because a fixed tick size prevents prices from neutralizing the effect of the maker rebate. A problem, however, with this extensive literature is that it does not take into account the switch in informed order flow away from a high liquidity subsidy exchange to its closest competitors when the subsidy is lowered. This may be caused by the change in the exchange s relative position of the routing table. From an empirical perspective, Malinova and Park (2015) analyze whether and why the breakdown of trading fees between liquidity demanders and suppliers matters by using a change in trading fees on the Toronto Stock Exchange (TSX), which is a monopoly exchange in their sample period. They find that posted quotes adjust after the change in fee composition, but that the transaction costs for liquidity demanders remain unaffected once fees are taken into account. In the absence of competition from competing exchanges, one does not expect any change in the composition of the order flow. However, we find that adverse selection costs decrease after the maker-taker fee reduces, holding the net exchange fee relatively constant 6. Colliard and Foucault (2012) theorize that in the absence of frictions, changes in the breakdown of the total fee between makers and takers should not affect trading behavior; only changes in the net exchange fee are economically meaningful for participant welfare. However, the informed traders in a competitive trading environment are not included in their model as patient and impatient traders are not informed traders. If an exchange introduces a maker rebate and finances it by adopting an increased taker fee 6 Because more than 90% of execution volume is displayed liquidity and more than 6% is non-displayed liquidity in our sample (Table 3), the overall net fee change remains relatively stable. 5

6 without changing the total fee, then, ceteris paribus, placing a market order becomes relatively more expensive than trading with a limit order. As some traders switch from market orders to limit orders, each limit order s execution probability declines and thus traders may improve quotes to attract matches for their limit orders. Without frictions, the benefit from maker rebates could be exactly offset by the narrowed bid-ask spread but only if the information content in the order flow is unaffected by the introduction of the matching fee subsidy. However, the implicit assumption of an unchanged order flow is implausible since informed trades with short-lived information need to be more aggressive to achieve a higher executive probability. Consequently, changes in the breakdown of the total fee between makers and takers should affect trading behavior and raise the spread, as liquidity trades are the first to be shifted toward making liquidity, while trades taking liquidity contain a higher proportion of asymmetric information. The alternative prediction of an unchanged spread made by Angel, Harris, and Spatt (2011) arises since they do not consider the altered order flow. While the Angel, Harris, and Spatt (2011) prediction is empirically supported by Malinova and Park (2015), this concurrence is most likely because they study the 2005 TSX fee change when it was a monopoly. In contrast, our study finds that when holding the total exchange fee relatively constant, the change in the breakdown of the fee components does matter. This difference may be driven by that our study examines the Nasdaq fee change in a competitive trading environment with 11 competing stock exchanges, approximately 40 active ATSs, and numerous broker-dealer platforms (Hatheway, Kwan, and Zheng, 2016). On a monopoly exchange, the variation in order flow informativeness is likely to be exceedingly small. Panayides, Rindi and Werner (2016) study the e 6 ffects of BATS Europe maker-taker fee reduction on market quality and market share in a fragmented market. Their model predicts, that a simultaneous reduction in make and take fees results in a deterioration of market quality and a significant spillover between venues following fee changes. However, they did not model the role of asymmetric information. This is important because when the rebates are reduced, it may force away the informed traders as the depth of the limit order book drops. Barclay, Kandel, and Marx (1998) empirically study how changes in bid-ask spreads influence volume and prices, finding that higher transaction costs reduce trading volume. Lutat (2010) argues that the Swiss Stock Exchange s removal of a maker fee (without changing the taker fee) has not affected the quoted spreads. Cardella, Hao, and Kalcheva (2015) study a number of maker-taker fee changes in the

7 United States from 2008 to They find that an exchange s total fee relative to that of other exchanges affects its trading volume and that a change in the taker fee has a stronger effect than a change in the maker rebate. They also show that the breakdown of the total fee into maker rebates and taker fees does not affect the quoted spreads. However, Malinova and Park (2015) argue that in Lutat (2010) and Cardella, Hao, and Kalcheva (2015), changes in the maker-taker fees are in fact accompanied by changes in the total fee. Foucault, Kadan, and Kandel (2013) examine the relative importance of the net fee and the levels of the maker and taker fees. They show that the breakdown of the total fee between makers and takers only becomes economically meaningful when the minimum tick size restricts adjustments to bid and ask prices but do not consider how the competition between maker-taker business models affects the information content of the order flow. Exchanges can maximize their trading volume by differentiating their maker rebates and taker fees. If liquidity demand (supply) is insufficient, the venue can decrease (increase) its taker fee and maker rebate. We also examine how such a maker-taker fee reduction affects market efficiency, which is an important aspect of market quality. The market efficiency hypothesis introduced by Fama (1970) emphasizes a lack of return predictability as the criterion for efficiency. Lo and MacKinlay (1988) tests market efficiency using variance ratio test which exploits the underlying property of the random walk process. Chordia, Roll, and Subrahmanyam (2008) present return predictability from order flows can arise in at least two ways. First, illiquidity raises implicit transaction costs (spreads) which squelch arbitrage activity. Second, market makers may mis-react to the information content of the order flow. This mispricing can create an incentive for outside agents to gather such information and trade on it. In this case, market makers will face increased adverse selection, which raises the equilibrium spread and leads to a less liquid market. Chordia, Roll, and Subrahmanyam (2008) find that liquidity enhances the incorporation of private information into prices. When the maker-taker fee is reduced, we typically find that the variance ratios increase, while first-order return autocorrelations decline in Nasdaq. This pattern suggests that the observed increase in these variance ratios is caused by increased mispricing, with less private information being reflected in prices. Moreover, compared with Nasdaq, the three other highest rebate-paying exchanges (Arca, BATS Z, NYSE) experience a relatively low decline in market efficiency. 7

8 Taker fees and maker rebates comprise a significant proportion of overall trading costs, given the typical bid-ask spread in a liquid stock is tick constrained, which is 1 cent set by the SEC for stock prices above $1. Although most retail investors do not have access to this information, all institutional investors and market makers, which account for majority of trading activity, make use of such data. Brokers commonly sell their marketable orders to wholesale dealers to capture the bid-ask spread and to avoid access fees, and send their non-marketable orders to exchanges for executions to gain maker rebate (Angel, Harris, and Spatt, 2015). In particular, they employ so-called smart order routers that take into account real-time state information and formulate an order routing problem that considers various execution metrics to decide whether to place a limit order or market order and accordingly to which venue(s) to direct their order (Maglaras, Moallemi, and Zheng, 2012). Because of its effect on execution probability in a fragmented market, we expect this liquidity impact of an exchange s fee-based net pricing to be increasingly important in the high frequency trading environment and perhaps explain the shift from lit exchanges to off-exchange venues. Battalio, Corwin, and Jennings (2015) examined the impact of differential exchange access fee schedules on broker routing decisions and limit order execution quality. They find maker-taker may be a mismatch between customer interest and broker incentive. Specifically, maker-taker may maximise liquidity rebate but result in poor execution quality for some clients. For nonmarketable orders, brokers may be incentivized to route customer orders to rest on a trading venue with the highest maker rebate. However, venues with high maker rebates are necessarily financed with higher taker fees, meaning that they tend to be ranked low on a broker s routing tables for routing marketable orders that take liquidity. Since the SEC Order Protection Rule does not specify who trades first when multiple trading venues have the best posted price, all else being equal, a price-sensitive market participant is more likely to route an order to take liquidity from an exchange with a lower taker fee. As a result, nonmarketable orders sent to rest on trading venues with high taker fees may experience lower fill rates (Battalio, Corwin, and Jennings, 2015). Brolley and Malinova (2012) argue that the effect of changes in the breakdown of the total fee into a maker rebate and a taker fee is not neutral if some traders (e.g., retail traders) only pay them on average, for example, through a flat commission to their brokers. In their model, only a fraction of traders receive maker rebates for each executed limit order. As the maker rebate increases, these traders improve their quotes and the raw bid-ask spread thus declines. Ceteris paribus, traders who pay the flat fee base their 8

9 order choices on the raw bid-ask spread rather than the cum fee bid-ask spread, thereby submitting relatively more market orders as the raw bid-ask spread declines. The authors then predict that in the presence of asymmetric information, the change in trader behavior causes market orders to become less informative. Their prediction is driven by the monotonic equilibrium behavior of traders in their model whereby, similar to Kaniel and Liu (2006) and Rosu (2012), traders with a sufficiently large informational advantage use market orders and those with weaker information use limit orders. This is due to privately informed traders facing a trade-off between the better price offered by a limit order and the potential loss of their informational advantage if their limit order does not execute. Those with the strongest informational advantage thus choose market orders and pay the bid-ask spread, whereas those with a weaker advantage choose limit orders and (hope to) receive the spread. However, HFT traders place either limit orders to gain rebates or market orders to pay a fee depending on their degree of information. In the maker-taker pricing model, the liquidity supplier has two main sources of revenue, namely bid-ask spreads and rebates, while liquidity takers profit from price movement minus the taker fee. In electronic markets, large market orders consume the available liquidity and widen the bid-ask spread. This drop in liquidity creates a profit opportunity for liquidity suppliers, which react by posting new quotes (make liquidity phase), which in turn create a new trading opportunity for liquidity demanders (take liquidity phase). Foucault, Kadan, and Kandel (2013) theorize this liquidity cycle process appears repeatedly. Each trading opportunity is short-lived as it disappears once a trader exploits it. Thus, HFT traders, who react faster than their competitors, gain the most from the profit opportunity. The SEC has requested comment on the HFT volume geared toward earning liquidity rebates and has questioned whether rebates [are] unfair to long-term investors because they necessarily will be paid primarily to [high-frequency] proprietary firms engaging in passive market making strategies. Or do they generally benefit long-term investors by promoting narrower spreads and more immediately accessible liquidity? (SEC, 2010). Angel, Harris, and Spatt (2011, p. 39) further argue that the makertaker model has aggravated agency problems among brokers and their clients because a typical broker does not forward the exchange fees to its clients on a trade-by-trade basis and may have a conflict of interests with its clients regarding the choice of trading venue. The strength of our analysis lies in the structure of the Nasdaq fee pilot change. First, Nasdaq introduced the maker-taker fee reduction only for a predefined subset of 14 securities, permitting the analysis of the 9

10 impact using a DID approach. Second, the pilot only lasted for four months in 2015, enabling us to compare PilotOn with PilotOff events. Third, Nasdaq data allow us to analyze how HFT reacts to the maker-taker fee reduction compared with non-hft across different resting order types, giving us new insights into the nature of HFT. Our results show that when Nasdaq reduces the maker-taker fee unilaterally without the cooperation of other exchanges, there are no significant changes in the market share of off-exchange trading venues (we classify these into dark pools and non-dark pools). Instead, we observe a redistribution effect among exchanges, whereby the reduced Nasdaq market share tends to shift to exchanges with higher maker rebates. Hence, holding the exchange net fee relatively constant, when the maker rebate reduces, the queue of Nasdaq limit orders shrinks, as evidenced by the drop in the percentage of time and depth at the National Best Bid and Offer (NBBO). Because the taker fee of removing liquidity declined (i.e., the tax on buyer- and seller-initiated trades fell), more market orders are submitted since they are less costly to trade. Moreover, the fill ratio and speed of fill increased as a wider class of investors could now profit with the lowered taker fee. Prior to the experimental pilot, the high taker fee discouraged all but relatively informed investors owing to the relative position of a market in routing tables. Since the exchange access fee is a material portion of the transaction cost given a stock is tick constrained, if one exchange s taker fee is relatively high, such a market tends to be ranked relatively low in the routing table. Thus, only the most informed orders are likely to reach it. Recall that traders possessing relatively short-lived information require immediate (liquidity taking) execution and hence are the most willing to pay the taker tax. With the lower access fee, Nasdaq would have moved up in the tables. Hence, we predict that the information content of trade will decline. The thinner limit order book deprived of its former subsidies becomes less attractive for informed traders and these are likely to switch to other exchanges that have retained their high liquidity subsidies in a competitive environment. Realized spreads should then increase to reflect the high but falling proportion of informed traders, but the cum-rebate (i.e., the net realized spread) decreases because of the smaller proportion of informed traders hitting limit orders. Hence, not only does the pilot reduce the net realized spread, but also the effective spread and market impact costs lower, which are all indicators of the switch in the informed order flow toward other high-rebate venues. As a net result, the Nasdaq market share declines on average and is captured by the other high rebate-paying exchanges in 10

11 close competition with Nasdaq. This finding indicates that the components of the exchange access fee do matter in a competitive trading environment. Since the provision of liquidity to highly informed traders is specialized, we also find a more than proportionate departure of HFT from the provision of liquidity but, more surprisingly, there is no net loss in HFT, as there is a shift to the expanding sector engaged in removing liquidity (i.e., market orders). One explanation, according to the SEC (2010) concept release, is the rebate capture by HFT; however, more relevant is the contribution of the speed of execution when dealing with highly informed order flows. The rest of the paper is organized as follows. Section II discusses the institutional details. Section III describes the data, sample selection, and research methodology. Section IV tests our empirical predictions on trading volume, market liquidity, and HFT trading behavior. Section V presents the results of our robustness tests and Section VI concludes. II. Institutional Details In this section, we describe the U.S. stock exchanges, off-exchange trading, and securities information processor (SIP) data. The following 11 lit exchanges account for approximately 65% of the market share in U.S. equities 7 : BATS Exchange (BATS), BATS Y Exchange (BATS Y), Chicago Stock Exchange (CHX), EDGA Exchange (EDGA), EDGX Exchange (EDGX), NASDAQ BX (BX), NASDAQ PHLX (PSX), Nasdaq Stock Market (NASDAQ), New York Stock Exchange (NYSE), NYSE MKT (AMEX), and NYSE Arca (ARCA). Among these 11 lit exchanges, NASDAQ, BX, and PSX are within the NASDAQ group (18.2% market share); NYSE, ARCA, and AMEX are within the Intercontinental Exchange group (24.6%); and BATS, BATS Y, EDGA, and EDGX are within BATS Global Markets (21.6% ) and Chicago Stock Exchange (0.6% of total market share) 8. At the exchange level, NASDAQ has the largest market share (approximately 15%) followed by NYSE and ARCA. NASDAQ follows the price, display type, and time execution priority model. 7 The CBOE Stock Exchange and National Stock Exchange, Inc. ceased market operations on April 30, 2014 and May 30, 2014, respectively. 8 Nasdaq Trader website in December 2015: 11

12 Off-exchange trades for NMS stocks, which account for about 35% of total volume, need to be reported to FINRA trade reporting facility (TRF) by members for which the Participant ID (Pid) is D in centralized SIP data. FINRA has established two TRFs in conjunction with NASDAQ and NYSE 9. U.S. equities trade execution venues can be classified into three main categories: exchanges, dark pools, nondark pools off-exchange venues which consists of ECNs 10, voice-brokered trades, and broker-dealer internalization. Dark pools and ECNs are also called ATSs. The primary difference between an ATS (typically operated by broker-dealers) and an exchange includes less regulatory scrutiny, fewer reporting requirements, and restricted access. III. Data, Sample Selection, and Methodology A. Data source Our analysis is based on trader-level data and U.S. SIP data. Our HFT data are identified by Nasdaq based on the method described in Brogaard, Hendershott, and Riordan (2014). We analyze the effect of Nasdaq s fee pilot for an eight-month window (December 1, 2014 to July 31, 2015), which is two months before to two months after the introduction of the maker-taker fee pilot, which ran from February 2, 2015 to May 31, We exclude the half-day trading on Christmas Eve. Our data include all information on order submission and trades, including price, volume, and a unique identifier for the trader that submitted the order, which allow us to construct the HFT data. We restrict our attention to transactions that occur in the limit order book and trades during regular trading hours. For each limit order book transaction, the data contain identifiers for buyer- or seller-initiated trade, adding or removing liquidity, and types of liquidity (such as displayed, non-displayed midpoint, and non-displayed non-midpoint). Further, our ATS data are provided via copyrighted by FINRA ATS data is reported weekly. Our 38 dark pools list is taken from the SEC report on the Regulation of NMS Stock Alternative Trading Systems. 9 FINRA Trade Reporting FAQ: 10 ECNs are prohibited from listing stocks and are not self-regulating organizations. The only remaining ECN in the United States, LavaFlow owned by Citi, ceased market operations on January 30,

13 B. Sample selection Nasdaq introduced a maker-taker fee reduction only for 14 stocks (seven listed on Nasdaq AAL, MU, FEYE, GPRO, GPRN, SIRI, and ZNGA and seven listed on the NYSE BAC, GE, KMI, RAD, RIG, S, and TWTR). We use the remaining companies on each corresponding listing exchange to find a oneto-one control group without replacement to ensure that our results are not driven by market-wide, exchange-wide, or industry-wide fluctuations. Moreover, we exclude securities that had stock splits, switched listing exchange, or had days with a stock price below $1. In our sample, the minimum trade count per stock per day is 450. Our control sample matches listing exchange, closing price, market capitalization, and average daily trading volume (ADV) based on one month prior our sample period data. Davies and Kim (2009) argue that one-to-one matching without replacement based on closing price and market capitalization is the most appropriate method to test for differences in trade execution costs. O Hara and Ye (2011) followed their approach and matched on closing price, market capitalization and listing exchange. We add ADV as a matching criterion since this study focuses not only on trade execution costs but also on trader behavior. In addition, we randomize the matching order by sorting the stocks in the treatment group alphabetically by ticker symbol. The match for each treatment group security i is then defined to be the control group security j that minimizes the following matching error: mmmmheeeee ii = CC i CC j CC i +CC j + MM i MM j MM i +MM j + AAA i AAA j AAA i +AAA j, where CP, MC, and ADV denote the security s closing price, market capitalization as of the end of November 2014, and average November 2014 ADV on its corresponding listing exchange (i.e., Nasdaq or NYSE), respectively. Our panel regression analysis employs a DID approach to account for marketwide fluctuations. The estimation is based on the following DID regression specification: Y ii = α 0 + α 1 TTTTT i + α 2 PPPPPPP t + β 1 TTTTT i PPPPPPP t + γvvv t + φx i + ε ii, (1) where Y ii is the dependent variable; α 0 is the intercept; TTTTT i is the dummy variable if security i is a pilot stock; PilotOn is the dummy variable that is one if date t is during the pilot period and zero otherwise; VVV t is the closing value of CBOE s volatility index for day t; and X i is the vector of security- 13

14 level control variables including the log of the average closing price during the sample period, the log of the average market capitalization during the sample period, and average volatility, measured by daily high price minus daily low price over closing price, during the sample period. We estimate the specification with and without stock fixed effects. Table 3 reports the summary statistics across U.S. equity trading venues for our sample of 28 stocks between December 1, 2014 and July 31, The sample period is divided into two parts: (i) the pilot period between February 2, 2015 and May 31, 2015 and (ii) the two months before and after the pilot from December 1, 2014 to February 1, 2015 and June 1, 2015 to July 31, The Nasdaq market share dropped by 1.45% on average during the pilot period, while the average price and market capitalization remained relatively stable. The Nasdaq depth at the NBBO and depth share declined, while the fill ratio and speed of fill improved. The raw effective spread increased, while the realized spread decreased, and the price impact decreased. By using Nasdaq data, we found that HFT changed trading behavior from adding liquidity (dropped by 10.04%) to taking liquidity (increased by 9.94%) after the reduction of the maker rebate and taker fee. The decline in HFT adding liquidity mainly came from the displayed liquidity type. <Insert Table 3 here> IV. Empirical Results A. Market quality and the information content of trades A1. How does the Nasdaq percentage of time at the NBBO, market depth, and fill rate change? Prediction: When Nasdaq reduces the rebate, the incentive to post liquidity is reduced, resulting in less quoting on Nasdaq at the NBBO. Thus, the Nasdaq percentage of time at the NBBO and corresponding depth and depth share will drop. After the taker fee is reduced, the fill rate and speed of fill is expected to increase. Result: After the maker rebate was reduced, the Nasdaq queue in the limit order book lowered. Table 4 shows that the Nasdaq percentage time and depth at the NBBO declined. The drop in the depth share at the NBBO (coefficient of *** in Table 4) is much larger compared with the slightly decrease in 14

15 market share (coefficient of *** in Table 7). This finding indicates Nasdaq s quote became relatively more attractive and their routing table position may have risen. Further, when the taker fee was reduced, the fill rate and speed of fill increased, since afterwards it was cheaper to remove liquidity by using market orders. <Insert Table 4 here> A2. How do the quoted spread, effective spread, realized spread, and price impact change? Prediction: Malinova and Park (2015) find that prices adjust as the maker rebate changes, using the TSX pricing change in In particular, holding the total fee constant, the raw bid-ask spread declines as the maker rebate increases. In addition, they show that the effective spread declines as the taker fee increases, consistent with the finding of Colliard and Foucault (2012). As the quoted spread declines, market orders become relatively cheap for these traders and more thus prefer market orders to limit orders. Brolley and Malinova (2012) show that this behavior is an equilibrium outcome, with the increased use of market orders causes a decline in adverse selection. The quoted spread is the difference between the ask price and bid price. Specifically, qqqqqad i,t = a i,t b i,t, (2) The effective spread is twice the signed difference between the transaction price and the midpoint of the bid and offer quotes at the time of the transaction. Specifically, eeeeeee i,t = 2 q i,t (p i,t m i,t ), (3) The realized spread is a measure of profit to market makers. Previous studies have set τ to five minutes after the trade. The choice of this time horizon should be sufficiently long to incorporate the permanent impact of the trade and thus to ensure that quotes are subsequently stabilized, temporary effects are dissipated, and there is a sufficiently long period for liquidity providers to close their positions (Conrad, Wahal, and Xiang, 2015). In today s ultra-high frequency trading environment that has upgraded trading systems with an accuracy of mere nanoseconds, five minutes is excessively long. Similar to Conrad, Wahal, and Xiang (2015), we estimate realized spreads from one second to five seconds after each trade. 15

16 The realized spread is then calculated as twice the signed difference between the transaction price and the midpoint of the bid and offer quotes one second and five seconds after the transaction. Specifically, rrrrrrr i,xx = 2 q i,t (p i,t m i,t+τ ), (4) Price impact is defined as the signed change between the midpoint of the quote one second and five seconds after the trade and the midpoint of the prevailing quote at the time of the trade. It captures the information that is revealed by the trade. A decline in the price impact indicates a decline in adverse selection costs. Specifically, ppppp iiiiii i,t = 2 q i,t (m i,t+τ m i,t ), (5) To test whether the fee change truly matters, we also compute the cum fee quoted spread, cum fee effective spread, and cum rebate realized spread as follows: ccc fff qqqqqqq i,t = a i,t b i,t + 2 f i,t, (6) ccc fff eeeeeee i,t = 2 q i,t p i,t m i,t + 2 f i,t, (7) ccc rrrrrr rrrrrrr i,xx = 2 q i,t m i,t+τ m i,t + 2 r i,t, (8) where f i,t is the taker fee for security i at time t, r i,t is the maker rebate for security i at time t, a i,t is the ask price of the quote, where b i,t is the bid price of the quote, p i,t is the transaction price for security i at time t, m i,t is the midpoint of the prevailing quote at the time of the trade, m i,t+τ is the midpoint of the one second and five seconds after the trade, and q i,t is an indicator variable that equals one if the trade is buyer-initiated and minus one if the trade is seller-initiated. Our data report the prevailing quotes as well as contain a marker that signs each trade as buyer- or seller-initiated. Result: Tables 5a and 5b shows the cum fee quoted spread and cum fee effective spread decrease (coefficients of *** and ***, respectively), which is inconsistent with the findings of Malinova and Park (2015). The cum fee quoted and effective spreads are adjusted by the taker fee, while the cum rebate realized spread is adjusted by the maker rebate. However, we find that the price impact 16

17 declined (-0.232*** for the one second price impact and *** for the five second price impact) after the maker-taker fee reduction. This decrease may have been caused by the relative position of a market in routing tables. In other words, if a market is ranked low in the routing table, only the most informed orders are likely to reach it. With a lower access fee, Nasdaq would have risen in the tables and depth reduced, resulting in a reduction in the price impact. Further, owing to the less competition in the limit order flow, the realized spread increases, while after taking into account the rebate reduction, the cum rebate spread decreases. Discussion The inconsistency between our spread and price impact findings and those of Malinova and Park (2015) 11 may be driven by the different trading environments analyzed. In 2005, TSX was the only stock exchange in Canada; however, our study of U.S. markets comprised 11 competing lit exchanges and many more off-exchange trading venues. In addition, liquidity suppliers have two main sources of revenue: the bid ask spread and rebate. Since traders had limited choices about where to trade in Canada in 2005, the impact rebate increase accompanied with increased fee had to be adjusted within TSX. Colliard and Foucault (2012) predict that the effective spread declines as the taker fee increases. Because the cost of market orders increases if the taker fee increase, traders switch to limit orders and the spread decreases. By contrast, in our competitive trading environment, when Nasdaq decreases the maker rebate, traders may post liquidity on other exchanges with a higher rebate. Hence, although the effective spread increases, we find a decreased cum fee effective spread because of the lower amount of information (i.e., lower toxicity) in the order flow as a result of the change in order routing. To compensate the liquidity supplier for the rebate loss, the realized spread must be widened for a given order flow type. After the rebate reduction is adjusted, the cum rebate realized spread decreases, indicating that the order flow is likely to be less informed. The high maker rebate financed with the high taker fee may attract informed traders given the other cheaper venues for removing liquidity, because the high rebate attracts a thick limit order book, which is particularly beneficial for informed traders. When the rebate is reduced, the limit order book gets thinner, which is less attractive to informed traders, although it is less costly to 11 The liquidity supplier does not necessarily need to adjust the realized spread since the trader has only the one trading venue to trade; hence, the realized spread does not change and the cum rebate realized spread increases. 17

18 remove liquidity via a market order. As a net effect, we find that the information content of the order flow declines. <Insert Tables 5a and 5b here> A3. How does market efficiency change? Prediction: To test the informational efficiency of prices, we use variance ratio tests of the random walk hypothesis (Lo and MacKinlay, 1988) and autocorrelation, which are identified as the two informationassociated measures in Chordia, Roll, and Subrahmanyam (2008), in multiple time intervals: VVVVVVVV rrrrr = vvv r i,t x/vvv(r i,x t ), (9) AAAAAAAAAAAAAAA = CCCC(r i,t, r i,t 1 ), (10) where vvv r i,t refers to the variance of the return during the t th time interval for i, vvv(r i,x t ) refers to the variance of the return during the x*t time interval for i and CCCC(r i,t, r i,t 1 ) refers to the autocorrelation of the midpoint return during the t th time period for i. Specifically, when a stock s price follows a random walk, the variance of its returns is a linear function of the measurement frequency, i.e. vvv(r i,x t ) is k times larger than vvv r i,t. When the maker-taker fee is reduced, we find that liquidity worsens (see the results presented above) and there is less information in the order flow. Thus, we expect the informational efficiency of prices to decrease. Result: Table 6 shows that in response to the maker-taker fee reduction, the variance ratios in Nasdaq generally increased, while first-order return autocorrelations declined. This pattern suggests that the variance ratios rise because of increased mispricing, with less private information being reflected in prices following the maker-taker fee reduction. Compared with Nasdaq, the market efficiency of the three other highest rebate-paying exchanges (Arca, BATS Z, NYSE) become less bad than Nasdaq given overall market efficiency declines. 18

19 <Insert Table 6 here> B. Trading volume and market share B1. Does a reduced exchange access fee attract trading volume from off-exchange venues? Prediction: The Nasdaq access fee pilot aimed to test whether a lower exchange access fee can raise the market share of off-exchange venues. In this regard, the SEC (2014) filing states the following: Offexchange orders do not generate quotes on public markets, do not interact with orders on public markets and consequently do not promote or contribute to price discovery to the same extent as do orders posted and executed on exchanges. Economic studies from markets spanning the world conclude that as more orders migrate away from exchanges, the price discovery process weakens, trading spreads widen, and overall investor trading costs increase Nasdaq believes that proposed changes may improve price discovery in the select securities. Kwan, Masulis, and McInish (2015) find that the U.S. minimum tick constrains some stock spreads, causing large limit order queues, and that dark pools allow traders to bypass existing limit order queues with minimal price improvement. Moreover, Foley and Putniņš (2014) find that when Canada and Australia implemented minimum price improvement, the level of dark trading decreased. We expect no change in dark pools trading volume after the maker-taker fee reduction since their ability to circumvent the time priority of displayed limit orders is not affected and nor are they required to provide price improvement. Result: Table 7 shows that after the Nasdaq fee reduction, the Nasdaq market share declined on average (see also Hatheway, 2015a, 2015b), while the consolidated trading volume remained stable and the share losses was captured by the other three highest rebate-paying stock exchanges (ARCA, BATS Z, NYSE). Table 8 shows that when Nasdaq reduced the maker rebate and taker fee, there was no significant market share drop from the off-exchange trading venues of dark pools and other non-dark pools. Instead, we observe a redistribution of market share among lit exchanges (i.e., the lost market share shifted to those lit exchanges with the highest liquidity provision rebate). One argument could be that only Nasdaq reduce the access fee unilaterally during the pilot during the competitive trading environment. Dark pools rely on mid-point prices and are protected by the inability of lit exchanges to reduce the minimum tick size because of the SEC s regulations. Instead, we observe a redistribution effect among existing lit exchanges. Even with a uniform fee reduction across the entire lit market, a movement toward the lit 19

20 market is unlikely as dark pools can still undercut the 1 cent tick rule and there is still no requirement to provide the minimum price improvement which, in any case, could be illusory. <Insert Table 7 here> <Insert Table 8 here> B2. How does the lit exchange trading volume and market share change in response to the maker-taker fee reduction? Prediction: In traditional maker-taker markets, the liquidity rebate, which is financed by a taker fee, is initially designed to increase liquidity. If the liquidity rebate is reduced, the trading volume and market share tends to decline. More specifically, when the rebate is reduced, the incentive to post limit orders on Nasdaq is reduced, the queue of limit orders declines, and the limit orders tend to flow to other exchanges with higher rebate payments. However, this effect is offset by the reduced taker fee since it is cheaper to remove liquidity as well. Overall, the Nasdaq market share declines since trade only occurs when there is liquidity supply, and the reduced volume goes to other exchanges with higher incentives. Result: Cardella, Hao, and Kalcheva (2015) find that reductions in relative taker fees in U.S. equity markets are associated with increased market share. Malinova and Park (2015) argue that the change in the maker-taker fees in Cardella, Hao, and Kalcheva (2015) are accompanied by changes in the total fee. Indeed, Malinova and Park (2015) find that holding the total exchange fee constant, TSX s trading volume in 2005 was unaffected by the component change. However, the DID regression model in Table 9 shows that the Nasdaq market share decreases in response to the maker-taker fee reduction. To verify whether this reduced market share is caused by a reduction in liquidity supply, we compare the Nasdaq depth at the NBBO change with the Nasdaq trading volume change. We find that the Nasdaq depth at the NBBO (coefficient of *** in Table 4) dropped about three times more than the Nasdaq trading volume drop (coefficient of *** in Table 9). We also did further statistical test and find statistically significant relationship between Nasdaq 20